Loss_scale = "dynamic" severely reduces training speed.
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 9k
- Forks
- 1.5k
- Avg merge
- 2d 4h
- Merged PRs (30d)
- 3
Description
I'm using apex for training a network on 4 GPUs. I'd like to use the O1 optimization level how ever it was actually slower than the fp32 training. I narrowed it down the the loss_scale being set to "dynamic" instead of a single float. Is this expected when I see my train step time reduced from 1.9s to 1s when using a float value vs "dynamic"? Is there something I can do to utilize dynamic loss scaling but still maintain the training speed similar to when I use a single float value?
Contributor guide
No contributing guide indexed for this repository
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start by reproducing the reported O1 training-step timings on four GPUs with a fixed loss scale and with dynamic loss scaling. Compare the runs and inspect the dynamic-scaling behavior to determine whether the slowdown is expected; document the result and any supported configuration that preserves training speed.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python, pytorch
- Domain
- machine-learning, performance
- Issue type
- Bug
- Difficulty
- 3/5
- Estimated time
- 1-2 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100